Noise, Cognitive Function, and Worker Productivity - Joshua T. Dean
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Noise, Cognitive Function, and Worker Productivity
Joshua T. Dean∗
January 22, 2019
Abstract
I investigate the relationship between noise and worker productivity with two exper-
iments in Kenya. I first randomize exposure to engine noise during a textile training
course. An increase of 10 dB reduces productivity by approximately 5%. I then ran-
domize engine noise during tests of cognitive function and effort. The same noise
change impairs cognitive function but not effort, suggesting the importance of cogni-
tive function as a mechanism. Finally, I demonstrate individuals neglect the effects
of noise by showing performance pay does not increase willingness to pay for quiet
working conditions. Individuals are thus unlikely to act to mitigate these effects.
∗
Institute on Behavior & Inequality; Address: briq, Schaumburg-Lippe-Strasse 5-9, 53111 Bonn, Ger-
many; Email: joshua.dean@chicagobooth.edu. I am extraordinarily grateful for the guidance and support
of my advisors Esther Duflo, Frank Schilbach and Abhijit Banerjee without whom this project would not
have been possible. This project also benefited from conversations with Alonso Bucarey, Emily Breza,
Sydnee Caldwell, Stefano DellaVigna, Dave Donaldson, Aicha Ben Dhia, John Firth, Chishio Furukawa,
Rachel Glennerster, Nick Hagerty, Johannes Haushofer, Peter Hull, Namrata Kala, Supreet Kaur, Gabriel
Kreindler, Matt Lowe, Benjamin Olken, Benjamin Marx, Rachael Meager, Matthew Rabin, Gautam Rao,
Otis Reid, Chris Roth, all of the participants at the Russell Sage Foundation Summer Institute in Behavioral
Economics, Joshua Schwartzstein, Mahvish Shaukat, Elizabeth Spelke, Carolyn Stein, Daniel Waldinger, Ro-
man Zarate and Ariel Zucker. I would also like to thank the entire team at the Busara Center for Behavioral
Economics for their superb advice and assistance with implementing this project. Finally, I would like to
thank the Busara Center for Behavioral Economics, the J-PAL Directors Fund, the George and Obi Shultz
Fund, National Science Foundation grant DGE-1122374, and the Weiss Family Program Fund for financially
supporting this work. This project was registered with the AEA RCT registry under ID AEARCTR-0001500.
1
While noise is ubiquitous in the modern world, in developing settings exposure can reach
egregious levels. In factories, workers experience jet-engine-level noise daily (Nandi and
Dhatrak 2008; Kimani 2011). On streets, it is common for honking and shouting to fill the air.
In India, this cacophony is so great that car manufactures have begun to increase the volume
of their horns (Sen Gupta 2014). Cognitive science research suggests that this exposure
might reduce productivity by impairing task management skills like attention and working
memory (Szalma and Hancock 2011; Matthews et al. 2000b; Jones and Broadbent 1991).
While it seems intuitive that inhibiting these cognitive functions should make individuals
less productive, no causal evidence demonstrates their importance for productivity. Without
this link, assessing the economic implications of excessive noise exposure is difficult.
This paper uses two randomized experiments in Kenya to investigate the relationship
between noise exposure and productivity. First, I estimate the reduced-form impact of noise
on productivity by randomly exposing participants in a textile training course to engine noise.
Second, I study the importance of cognitive function as a mechanism by randomly exposing
individuals from the same population to the same engine noise while they complete a battery
of cognitive tests.1 Third, in both experiments I assess whether individuals understand
how noise affects their productivity by offering participants the chance to pay for quiet
working conditions while I randomly vary whether their pay depends on their performance.
Understanding whether individuals appreciate the impacts of noise on their productivity is
necessary to draw conclusions about welfare. If individuals completely understand the impact
of noise, we might expect the effects to be attenuated by adaptations such as avoiding noisy
locations.
I demonstrate that noise can meaningfully reduce productivity in a real-work setting.
While a significant body of literature considers the impact of noise on cognitive function,
very little work considers how this effect might manifest itself when individuals are faced with
incentivized tasks in a real-work setting (Matthews et al. 2000b). I recruited a sample of 128
manual laborers accustomed to factory noise for a two-week textile production course at a
vocational training facility outside of Nairobi, Kenya. After training the sample to complete
an incentivized production task, I randomly exposed participants to engine noise while they
worked autonomously for a piece rate. In order to isolate the impact of noise, I chose a
task that did not involve communication, randomly varied which work location was noisy,
and randomized work stations to minimize participants’ familiarity with their neighbors.
I estimate that increasing the noise level from that of a dishwasher to that of a vacuum
1
This mechanism holds particular interest because a recent literature in economics considers how condi-
tions of poverty might affect cognitive function, but no work has yet estimated a causal relationship between
cognitive function and economic activity (Mani et al. 2013; Schilbach et al. 2016).
2
cleaner (an increase of 10 dB) reduced output by approximately 5%. For comparison, Bloom
et al. (2013) find that an intensive, nine-month management intervention increased output
by 9%.
Given that the task did not involve communication, the most plausible channel for this
impact is through the effect that noise has on cognitive function. Cognitive function (also
called executive function) encompasses all of the general-purpose abilities involved in task
management. This includes the ability to direct one’s attention, manipulate information in
memory and switch between tasks (Diamond 2013). These skills appear critical for many
types of work. For example, a factory foreman requires a broad range of attention to ensure
that his/her workers do not make mistakes. An auto-rickshaw driver must simultaneously
drive and take directions from his/her passenger. Cognitive science research has shown
stronger cognitive function is correlated with better job market outcomes, physical health
and success in school (Bailey 2007; Borella et al. 2010; Crescioni et al. 2011; Duncan et
al. 2007; Gathercole et al. 2004). Nonetheless, we do not have any causal estimates of the
impact of these task management abilities on real-world economic outcomes.
Decades of cognitive science research has highlighted the damaging effect that excessive
noise exposure has on cognitive function (Evans and Hygge 2007; Hockey 1970; Jones and
Broadbent 1991; Matthews et al. 2000b; Smith 1989; Szalma and Hancock 2011). Noise has
been shown to inhibit performance on tasks that require broad attention, management of
several component sub-tasks and memory recall, especially the loud and variable types of
noise common in developing contexts (Hygge et al. 2003; Irgens-Hansen et al. 2015; Jahncke
et al. 2011; Kjellberg et al. 2008). This large body of literature suggests that noise pollu-
tion provides a promising context in which to study the importance of cognitive function
for real-work tasks. It provides both assurance that cognitive function is likely an impor-
tant mechanism and guidance on what alternative channels to examine, namely reduced
motivation and stress-induced physical impairment.
To evaluate the importance of this mechanism, I randomly exposed individuals from
the same population to noise while they completed a wide variety of cognitive tests. The
same engine noise reduced performance on a common factor index of test outcomes by 0.07
standard deviations. In order to assess alternatives, I also recorded respondents’ blood
pressure and asked them to complete an effort task (DellaVigna and Pope 2018). While the
increased noise did lead to a small increase in blood pressure, suggesting that respondents
experienced more stress, the effect is too small to have had physical effects. Moreover, I
find no change in performance on the effort task. Combined, this evidence suggests that
physical distress or diminished motivation are unlikely to be important mechanisms. Of
course, there could be alternative channels not considered in the literature, but if we are
3willing to assume that cognitive function is the primary mechanism, the combined results
of these two experiments imply that cognitive function is an important input to worker
productivity.
I then demonstrate that individuals neglect the productive impact of noise. While a
significant literature examines the disutility individuals derive from living in noisy conditions
(see Navrud (2002) for an overview), no work has assessed whether individuals are aware of its
productive impact. More generally, while a rapidly expanding literature suggests experiences
associated with poverty might have important economic consequences by impairing cognitive
function, almost nothing is known about whether individuals take actions to protect their
cognition from these stimuli (Kremer et al. 2019).2 3 Understanding whether individuals are
aware of environmental hazards to their productivity is critical to predicting the real-world
impact of such impediments. If individuals understand the effects and have the ability to
either avoid or compensate for exposure to these stimuli, we would expect their actions to
attenuate any effect observed in a controlled experiment. If, on the other hand, individuals
do not appreciate their vulnerability, these environments are potential sources of significant
inefficiencies.
In order to assess this possibility, I allowed participants in both experiments to pay
for quiet working conditions and randomly varied whether they were paid based on their
performance. If individuals attend to the productive effects of noise exposure, they should
be willing to pay more to work in quiet when their earnings depend on their performance.
Instead, I find that individuals’ willingness to pay was unaffected by the wage structure.
I use my within-person variation to evaluate potential mechanisms underlying this ne-
glect. I first assess whether individuals who were relatively unaffected by noise are driving
the result. By estimating individual-level treatment effects, I show that the impact of noise
on an individual’s productivity does not predict the responsiveness of their demand to per-
formance pay. Second, I show that a simple prompt to think about the productive impacts
of noise also did not increase demand responsiveness. Finally, I demonstrate that responses
are consistent with a failure to notice the productive impact of noise (Hanna et al. 2014;
Schwartzstein 2014). In particular, individuals were able to somewhat predict their output
but were unable to predict the impact of noise. Moreover, individuals appear to have realized
that they did not understand the productive value of quiet and were unwilling to stake any
2
For example, there is recent work on on effects of alcohol consumption, heat, and air pollution (Schilbach
2017; Adhvaryu et al. 2016; Zivin and Neidell 2012), and on-going work on studying the effects of sleep
deprivation, pain, and resource scarcity.
3
Schofield (2014) demonstrates that individuals fail to make food purchases that would improve their
productivity which suggests possible neglect. Also previous work in psychology on the human capacity for
introspection (for example, Nisbett and Wilson (1977)) also suggests individuals in these environments might
not be aware of their potential impact.
4money on their stated beliefs. I conclude by considering implications of this neglect for noisy
firms’ ability to efficiently screen affected workers.
The remainder of this paper is organized as follows. Section 1 discusses the prevalence
of noise pollution in developing cities and its effects on cognitive function, before Section 2
describes the design and results of the productivity experiment. Section 3 then presents the
design and results of the cognitive experiment, Section 4 assesses whether individuals neglect
the effects of noise and considers implications for efficiency, and finally Section 5 concludes.
1 Background
1.1 Noise Pollution in Developing Contexts
Noise pollution is one of the oldest externalities documented in the written record. In the
6th century BCE, the Greek colony of Sybaris had such a noise problem that they banned
potters, tinsmiths and other noisy tradesmen from working in the city (Goldsmith 2012).
When the founding fathers of the United States gathered in the Pennsylvania State House
in May 1787 to craft the constitution, they first spread dirt on the cobblestone streets
surrounding the building to prevent the noise of passing carriages from disrupting their work
(United States National Archives and Records Administration 2017). Since the industrial
revolution, sources of noise pollution have proliferated at an impressive rate (Bronzaft 2002).
Given weak state capacity, it is unsurprising that noise pollution is pervasive in the rapidly
urbanizing and industrializing developing world. In many cities the noise level experienced
by simply standing on the street reaches dangerous levels (Wawa and Mulaku 2015; Mehdi
et al. 2011; Bhosale et al. 2010). For example, areas of the central business district of Nairobi
approach 85 dB (the level of noise made by a power lawn mower).
Beyond city streets, many workplaces are filled with noise. The Indian National Institute
of Occupational Health reports that noise levels in most industrial occupations exceed 90
dB, a level that the United States Centers for Disease Control estimates will induce disabling
hearing loss in one out of four workers exposed (Nandi and Dhatrak 2008). Similarly, an
NGO in Kenya finds that 75% of metal workers are exposed to unsafe levels of noise and
22% already have disabling hearing loss (Operation Ear Drop 2010).
While comprehensive data on noise levels does not exist outside of the European Union,
we can use hearing loss as a proxy for exposure. Figure 1 combines measurements of hear-
ing ability recently collected by Mimi (2017) with data on city-level GDP from Berube et
al. (2014) to show that citizens of poorer cities have substantially more age-adjusted hearing
loss. The average citizen of Delhi or Mumbai has as much hearing loss as residents of New
5York or Tokyo who are eight years older.
1.2 Noise and Productivity
Despite the research on the cognitive impacts of noise, we have almost no causal evidence of
the impact of noise on economic outcomes in real-work settings. Weston and Adams (1935)
randomized hearing protection among 20 textile workers and estimated that output was 3%
higher among those with hearing protection over the next 18 months. Unfortunately, the
study does not report standard errors or any statistical tests which makes it difficult to
interpret this result. Broadbent and Little (1960) studied the effects of a noise-abatement
treatment in one room of a Kodak factory and found that the noise decrease of 10 dB
was associated with fewer worker errors; although, there was also a significant improvement
in the non-abated rooms. Finally, Levy-Leboyer (1989) cross-randomized 52 workers into
assembling either carburetors or air conditioners in either their typical noisy conditions or a
separate quiet room. Workers assigned to assemble air conditioners in quiet were 14% faster
than those in normal conditions; however, those assigned to assemble carburetors in quiet
were 10% slower than their counterparts in noise. Although no study provides large-sample
evidence that distinguishes the effects of noise exposure from other location-specific features,
together this work suggests that noise might affect real-work outcomes.4
1.3 Cognitive Function and Productivity
Studies on cognitive function and productivity generally fall into one of two groups, the first
of which examines how stimuli can affect cognitive function. A large psychology literature
studies how a variety of factors such as heat, fatigue, sleep, and health can affect cognitive
performance (see Matthews et al. (2000a) or Dean et al. (2017) for overviews). Additionally,
recent literature in economics examines how conditions of poverty can impede cognitive
function (Haushofer and Fehr 2014; Lichand and Mani 2016; Mani et al. 2013; Schilbach et
al. 2016). These studies then generally appeal to theory, the correlational evidence mentioned
above, and our intuition about the importance of cognitive abilities to make inferences about
how stimuli might affect productivity.
A second group of studies examines how stimuli can affect productivity directly. This in-
cludes recent work in economics on how temperature, alcohol, air pollution, and hunger can
affect productivity (Adhvaryu et al. 2016; Chang et al. 2016b, 2016a; Park 2017; Schilbach
2017; Schofield 2014; Zivin and Neidell 2012). While these studies provide invaluable ev-
4
Researchers have studied the impact of OSHA regulations on productivity; however, such work is unable
to separate the effects of noise regulations from other safety regulations (Denison 1978; Gray 1987).
6idence on the potential for environments to affect productivity, they are unable to speak
directly to the importance of a cognitive mechanism because the factors that they study
generally affect productivity through multiple channels.
To my knowledge, no work bridges the gap between these two groups and studies how a
single stimulus affects productivity through cognitive function. Without tracing this entire
path, interpreting the importance of environmental stimuli where we only have evidence
on cognitive effects is difficult. In particular, in order to evaluate the importance of the
cognitive impediments associated with poverty, we need an estimate that translates these
effects into economic outcomes (Schilbach et al. 2016). As noted above, the large literature
on the cognitive effects of noise exposure makes an exploration of this cognitive channel
possible.
Finally, almost none of this work directly evaluates whether individuals understand how
environments can affect their productivity via cognitive impediments (Kremer et al. 2019).
If individuals appreciate these impacts, they might be able to take actions that significantly
attenuate the effects estimated in controlled experiments. However, there is some reason to
doubt this is the case. Schofield (2014) finds individuals fail to consume calories that would
improve their productivity, Adhvaryu et al. (2016) reports that managers were surprised
by the results of their study demonstrating heat reduced productivity, and research in psy-
chology such as Nisbett and Wilson (1977) suggests individuals may not have the capacity
to truly monitor their own cognitive processes. This study provides direct evidence on this
question by demonstrating demand for quiet working conditions is unresponsive to incentives
to increase productivity.
2 Experiment One: Noise and Worker Productivity
This experiment provides reduced-form evidence of the impact of noise on productivity.
By randomly exposing workers in a textile training course to engine noise, I estimate that
increasing the noise level by 10 dB (from the noise level of a dishwasher to that of a vacuum)
reduces output by approximately 5%.
2.1 Experimental Design
2.1.1 Context
My survey team recruited 128 individuals for a ten-day sewing course at the Kenyan Na-
tional Industrial Training Authority’s Technology Development Center (TDC), a vocational
training facility located in an industrial development zone outside of Nairobi. We recruited
7our sample from groups of manual laborers who gather at the gates of nearby textile factories
hoping to be hired for a day’s work (see Figure A1).
This population is well suited for this experiment for three reasons. First, the fact
that respondents typically work in factories means that they are accustomed to significant
levels of noise. Second, these participants have the opportunity to use the skills learned
in the course to gain employment, which helps the experience approximate typical working
conditions. Third, the sample is demographically similar to many poor communities where
we are interested in the importance of cognitive function (Table B1).
2.1.2 Generating Noise
There are three ways to manipulate noise exposure: ambient-level abatement, individual-
level protection, and generating additional noise. Ambient-level abatement is undesirable
from an experimental perspective because it involves significant, location-specific invest-
ments that confound the reduced noise with other location-specific features. For example, a
common abatement technology is to replace or pad the existing ceiling with more absorbent
material. While effective at reducing noise, this means that those randomized to the room
with the absorbent ceiling are necessarily also treated with the other features of that room
such as temperature, humidity, and ventilation. Individual-level protection does not involve
location-specific investments, but noise control experts view it as an option of last resort
due to its relative ineffectiveness and the safety risks that it creates by impeding workers’
ability to warn each other about hazards (Hansen and Goelzer 2001). Additionally, hearing
protection not only alters the experienced noise level but also affects the physical comfort
of the participants. For these reasons, I chose to manipulate noise by adding a new noise
source to the preexisting noise generated by the sewing machines.
In order to create noise representative in both the level and quality of that faced by
residents of developing countries, I chose to generate noise with a car engine that the TDC
uses for auto-mechanic training classes (see Figure A2). This type of engine noise is repre-
sentative of both noise pollution generated by traffic and occupational noise generated by
large industrial machines. This has two benefits: first, the effect of noise is known to depend
on predictability and variability (Matthews et al. 2000b), thus the representative nature of
the noise is important for external validity; and second, this type of noise is unlikely to be
novel to participants, which limits concerns about whether any productivity effects are due
to respondents simply changing behavior in response to a novel stimulus. The end result
is that participants in the control condition experienced noise approximately equal to that
of a home dishwasher running in the background, while in the treatment condition workers
experienced noise equivalent to listening to a home vacuum cleaner.
8One might be concerned that in addition to creating noise, engine exposure could alter
other environmental conditions. For example, engine exhaust might diminish the room’s air
quality or annoyance with the noise might cause participants to close windows, changing
the temperature inside the room. These altered environmental conditions could then have a
direct effect on productivity independent of any effects of the noise level. Thus, in order to
ensure that treatment only increased noise exposure, enumerators were instructed to keep
the windows and doors unchanged and ensure that the exhaust pipe from the engine pointed
away from the workroom doors into an open courtyard. To assess whether this was successful,
I measured CO2 (as a proxy for engine exhaust), temperature, and humidity during every
session.5
2.2 Production Task
I chose sewing pockets as the incentivized production task for several reasons. First, it is
a task that can be completed relatively quickly, which allowed me to observe variation in
performance over a short time period. Second, it requires many key sewing skills (e.g. sewing
under control, sewing in parallel lines, hemming, and taking corners). In fact, the TDC uses
this task as a tool to evaluate potential instructors for precisely these reasons. Third, these
sewing skills in turn require a variety of cognitive functions. For example, sewing in a straight
line requires paying close attention to how hard one presses the machine foot pedal, how
quickly one moves the fabric with both hands, and exactly where the needle is puncturing the
fabric being sewn at all times. These cognitive requirements are common to many production
tasks that workers perform in developing contexts, which improves the external validity of
the study. Third, this task does not require communication. If the task I chose for the study
required participants to communicate, any observed effects would be the result of impairing
both communication and cognitive function. This would then preclude me from using this
experiment to explore the importance of cognitive function as a mechanism and participants’
awareness of this importance. Finally, the task does not generate considerable noise. If the
task I chose created significant noise (for example metal work), I would observe a mechanical
positive correlation between noise and productivity.
The quality of the pockets produced was graded each hour by treatment-blind enumer-
ators according to six criteria developed by the TDC (see Figure 3 for an example pocket
with the criteria marked). In the analysis below, I use these grading data to construct three
types of productivity measures. First, I use the number of pockets created per session as
a pure quantity metric. Second, I combine quantity and quality by calculating the number
5
CO2 is typically highly correlated with other exhaust pollutants such as particulate matter and black
carbon (Johnson et al. 2016; Abdel-Salam 2015).
9of “points” earned across all pockets produced in a session. For example, if a subject made
one pocket meeting four criteria and another meeting three criteria, they would earn a total
of seven points. This is my most continuous metric where I have the most power. Finally,
I report the number of pockets meeting the different possible quality thresholds per session.
For example, the number of pockets meeting two criteria, the number of pockets meeting
three criteria, and so on. The distribution of these outcomes is skewed, but has zeros (see
Figure A4). Thus, in order to improve power I use inverse hyperbolic sine transformations
as my preferred outcomes following Burbidge et al. (1988).6 For robustness I also present
the results in levels. All of the outcomes yield similar results.
2.2.1 Experiment Timing Overview
For logistical reasons, the course was repeated in four rounds with the number of respon-
dents equally split over each round. On the first two days of the course, TDC staff taught
participants how to operate a sewing machine (see Figure 2a). This included basic skills such
as how to thread the machine and how to avoid breaking the sewing needle. After learning
these basic skills, workers then learned how to sew a pocket. All training occurred without
engine noise.
Respondents worked three sessions per day for the remainder of the experiment, sewing
pockets and earning a piece rate for each perfect pocket that they created. On the last two
days, respondents had the opportunity to pay to work in quiet. On all days, participants
worked for three two-hour sessions separated by one-hour breaks without knowing their
future treatment status (see Figure 2b). These breaks allowed workers who were in the more
noisy environment to decompress between sessions. Combined with the lack of knowledge
about future treatment status, this allows me to isolate the contemporaneous effects of noise.
This improves my power in the analysis below because it allows for the pooling of all workers
within a session based on their contemporaneous treatment status, rather than having to
include interactions with their previous or future exposures.
2.2.2 Lasting Effects of Noise
While I designed this schedule to isolate the contemporaneous effects of noise, whether noise
exposure has lasting effects is an important policy question. I thus also include the following
decision tasks that were completed in quiet at the end of the day:
p
6
An inverse hyperbolic sine transformation is defined as f (y) = ln y + 1 + y 2 . It has the benefit
that, except for values of y close to zero, f (y) ≈ ln(2) + ln(y). Thus, as long as there are not too many
zeros and values are reasonably large, coefficients can be interpreted in a similar manner to a standard log
transformation.
101. On every production day, participants decided how much to save in/withdraw from
an account with a 1% per working-day interest rate (approximately 7% interest over
the course of the experiment). This was intended to assess whether noise exposure
reduced willingness to forgo current consumption by either raising the contemporaneous
marginal utility of consumption or narrowing attention to the present.
2. On the fifth day, participants decided whether to buy maize flower in 5 kg bags or 1
kg bags. To test for increased inattention prices were set so that it was less expensive
to buy five 1 kg bags than one 5 kg bag.
3. On the sixth day, participants decided whether to stay an additional hour and con-
tinue working for a piece rate. This was intended to assess whether noise reduced
participants’ willingness to exert effort.
2.2.3 Randomization
For each of the sessions following training, I randomized which participants were exposed to
engine noise while working. For this purpose, I generated random schedules for each round
that satisfied both of the following constraints:
• Each worker spent half of the time in noise and half in quiet.
• In each session, an equal number of participants worked in quiet and in noise.
I then randomly assigned each participant to one of the schedules. In order to avoid any
location-specific confounds, participants worked in two similar rooms, and I randomized
which room was noisy for each session7 (see Table B2 for balance tests). This generalization
of stratification was necessary because my piloting demonstrated significant heterogeneity
in the ability to complete the production task across both individuals and time. Thus, even
though simple randomization procedures would have resulted in balance in expectation, the
risk of imbalance in finite samples was substantial. Following this randomization method, I
include worker, room and session fixed effects in my regressions, which significantly improves
my power.
2.2.4 Compensation
For each production session, I randomized workers to one of three wage conditions. Each
wage was a combination of a piece rate paid based on the number of perfect pockets produced
7
The rooms were located within walking distance of each other in the compound, but not so close that
sound could travel from one to the other.
11and a flat payment for participation. All three conditions were calibrated to yield approx-
imately 200 Ksh per session (or 600 Ksh per day), although they differed in the relative
importance of the piece rate and flat payment.8 This allows me to benchmark the observed
effect of noise against the effect of traditional incentives.
2.3 Analysis and Results
2.3.1 Environmental Effects of Treatment
Treatment increased the noise level by approximately 7 dB (Figure A3 and Table 1).9 As
noted previously, this difference is equivalent to the difference in noise between a home
dishwasher and a home vacuum cleaner. Meanwhile, no other environmental variables were
affected, suggesting that the pollution and temperature control procedures were effective.
2.3.2 Estimation Specifications
I estimate two different specifications. The first is the reduced-form effect of being in a
treated room on productivity outcomes for individual i in room j at time t being paid wage
w shown in equation (1). The regression includes individual, time, room, and wage fixed
effects and has standard errors clustered at the level of randomization (room × session).
yijtw = τ · Treatmentjt + αi + γt + φj + κw + ijtw (1)
yijtw = ν · Noise Leveljt + αi + γt + φj + κw + ijtw (2)
To improve interpretability, I also estimate an instrumental-variables specification shown in
equation (2) using an indicator for being in a treated room as an instrument.10
8
On training days, all respondents received 600 Ksh (approximately $6.00) for participating. The produc-
tion day piece rates were one of 15, 10, or 5 Ksh per perfect pocket. In the first round, the corresponding flat
rates were 140, 160, and 180 Ksh, respectively. After participants in the first round were more productive
than anticipated, the flat rates were reduced to 95, 130, and 165 Ksh to make the wage treatments as income
neutral as possible. All wage fixed effects are determined based on the piece rate, which is common across
all rounds.
9
For interpretability, all noise levels are reported in 10s of decibels because the human ear perceives an
increase of 10 dB as a doubling of the noise level. Thus, coefficients can be interpreted as the effect of
doubling the noise level.
10
Using only a simple indicator for treatment discards the significant variation in treatment intensity
shown in Figure 5. Since this treatment intensity is quasi-randomly determined based on the noise levels at
the compound and whether the engine was running smoothly or rattling, in Appendix B I use this variation
to obtain more precise estimates by generating separate treatment indicators for each decile of intensity
(difference in noise level between treatment and control room), Treatmentjpt , that are equal to one if room
j was treated during a session with intensity p and zero otherwise. There are no clear relationships between
session intensity and any observable characteristics besides noise (see Table B5) and the instruments yield a
strong first stage (see Table B6).
122.3.3 Main Results
Workers sewing in treated rooms produced approximately 3% fewer pockets (Table 2). Scal-
ing this by the average noise change implies a 5% decrease in productivity for every 10 dB
increase (or perceived doubling) in the noise level (Table 3).11 In these specifications, there
appears to be no effect of the noise on the number of perfect pockets. This is likely due
to floor effects. During the first days of work, most participants were unable to make any
perfect pockets. Because individuals cannot produce negative pockets, this attenuates the
estimated treatment effect.12 Fisher p-values shown in Table B3 yield similar inferences.
These effects are unlikely to diminish with further exposure to noise. As noted above,
the participants are already accustomed to working in large, noisy factories, and they are
exposed to frequent road noise (the community sits at the intersection of two major highways
from Nairobi to Mombassa and Arusha). What is less certain is how these effects map into
different types of tasks. These sewing tasks were chosen explicitly because they appear to
depend on cognitive function. It is unlikely that noise exposure would impede the ability
of someone doing a less cognitively demanding task such as selling water on the side of the
road. On the other hand, many factory employees are required to work in teams assembling
complex objects, and noise would likely impede both each individual’s cognitive function
and the team’s ability to coordinate. Another complication with extrapolating from these
effects is that different sources of noise pollution vary in predictability and informational
content. While the noise in this study was chosen to be representative in level and quality
of major sources of noise pollution, they are by no means the only sources. Further research
is needed to understand the effects of other common sources, such as your child overhearing
your neighbor’s television while trying to study.
2.3.4 Treatment Effect Heterogeneity
One might wonder whether this effect is driven by low-ability workers. If this is true, then
a firm could eliminate the effect by firing the bad workers. To assess this possibility, I
calculate each individual’s performance in the control condition and split the sample at
11
Using a larger portion of the variation by generating separate instruments for different treatment inten-
sities yields slightly larger coefficients (see Table B7).
12
An alternative explanation is that workers substitute their effort so that they make fewer but higher
quality pockets. This is not borne out in the data. If effort substitution were occurring, we would expect
the proportion of perfect pockets to be higher in treatment than in control. Table B8 shows this is not
the case. Additionally, estimates of the reduced-form in the second week, although imprecise, show a larger
effect on perfect pockets than on total pockets, suggesting that once individuals are capable of producing
perfect pockets their production is also affected by noise exposure (Table B4).
13the median.13 I then estimate the treatment effect separately for each group in a stacked
regression with common fixed effects.14 The treatment effects are equally large among better
workers (Figure A5).
Another question is whether these results are specific to contexts in which individuals
are learning how to perform a task or where they have not yet had the chance to adjust fully
to their surroundings. I assess this possibility by estimating the treatment effect separately
for each week and for the first and second hours of each session. While these regressions
have lower power, the effects appear constant across both dimensions (see Figure A6 and
Figure A7).
2.4 Lasting Effects of Noise
Some models of the effects of noise predict that exposure should generate lasting effects
(Matthews et al. 2000b). While this study was designed to minimize these effects by in-
cluding breaks in the schedule, I assess whether the effect of noise exposure is cumulative
by regressing the inverse hyperbolic sine-transformed outcomes on a treatment indicator, a
lagged treatment indicator and their interaction.15 The results are imprecise but do not sug-
gest that lagged exposure is important (Table B10). Additionally, treatment did not affect
any of the decision tasks (Table B11). This seems to suggest that the effects of noise do
not persist into later periods of quiet; however, given that the experiment was not primarily
focused on this question and power in these tests is low, future work should evaluate this
question directly.
2.5 Conclusions from Experiment One
These results suggest that noise pollution has the potential to have a significant impact on
productivity. In their noise survey of central Nairobi, Wawa and Mulaku (2015) found that
noise levels ranged between approximately 55 dB (the level of a background conversation)
and 85 dB (the level of a lawn mower). If we believe that the effect of noise remains linear
outside of the range considered in this experiment, this implies that workers in the quieter
areas of Nairobi are 15% more productive than those in the more noisy areas.
To interpret these magnitudes, it is helpful to compare them to other methods of improv-
ing productivity. In this experiment, doubling the piece rate from 5 Ksh to 10 Ksh while
13
I exclude the current session from the calculation to avoid the overfitting problems highlighted by Abadie
et al. (2014).
14
This procedure is equivalent to first partialling out the fixed effects and running the regressions separately.
15
The lagged treatment indicator is set to zero for the first session of each day.
14lowering the flat rate to compensate raised output by 3%.16 Kaur et al. (2015) found that
offering commitment contracts increased output by 2.3%. Finally, Bloom et al. (2013) found
that a five-month intensive management intervention in an Indian textile firm increased
output by 9%.
Another way to interpret the size of these effects is to consider how these estimates might
affect firm noise-abatement decisions. Unfortunately, it is impossible to make any general
claim about cost-effectiveness because abatement costs are highly context-specific. Costs
can vary by orders of magnitude depending on the noise’s source, the building structure,
and the production processes (Hansen and Goelzer 2001). Nevertheless, one can consider
whether this effect is sufficiently large to be relevant to some firms’ abatement decisions.
In particular, Lahiri et al. (2011) report a case study of a large computer manufacturing
firm in Singapore, where reducing the noise level by 23 dB cost the firm $156 per worker
per year. Combining this cost with my estimate and assuming all productivity gains would
translate into increased profits implies that this firm would break even on abatement if each
worker produced $1,357 per year ($5 per day) in profit. For comparison, workers at this
firm were paid $12.50 an hour. This suggests that, at least for some firms, a 5% increase in
productivity is sufficient to affect abatement decisions.
Additionally, these effects are as large or larger than other environmental pollutants
studied in the literature (Table B9). This suggests that as policy makers consider priorities
in managing the explosive urbanization and industrialization of the developing world, they
should not neglect noise pollution. Even simple regulations such as limiting the volume of
car horns can prevent a race to the bottom that imposes costly externalities.
3 Experiment Two: Noise and Cognitive Function
While the evidence from experiment one demonstrates the potential importance of noise
pollution, it does not speak to the underlying cognitive mechanisms. These mechanisms
hold particular interest because recent work argues that conditions of poverty may trap
participants by impeding their cognitive function (Schilbach et al. 2016). However, without
causal evidence on the importance of these cognitive functions for economic activity it is
difficult to evaluate the plausibility of this claim.
I use this second experiment to provide the some of the first evidence on the importance
of this mechanism in three steps. First, I estimate the effect that the same noise change has
16
Increasing the piece rate from 5 Ksh to 15 Ksh had no effect on output. One explanation consistent with
this evidence is that even though the flat rates were calibrated to compensate on average, income effects
began to mitigate the piece rate’s effectiveness as an incentive.
15on cognitive function. Second, I evaluate whether noise affected the alternative mechanisms
highlighted in the extensive psychology literature (namely changing the technology of the
task, reducing motivation and causing physical distress). Finally, because I only find evidence
of an effect on cognitive function, I use a split-sample IV to compute a hypothetical “return to
cognitive function” to illustrate how the combined evidence from these experiments implies
cognitive function is an important input to productivity.
3.1 Design Overview
In order to examine the mechanisms in a credible way, I replicated the conditions of the first
experiment as closely as possible. I used the exact same recruiting procedure (see Table B1
for a comparison of sample demographics). The experiment was conducted in similar rooms
less than a mile away from the TDC (see Figure A1 for a map of the locations). The timing
was a condensed version of experiment one (see Figure 4). Participants came for two-day
rounds. They first spent two hours learning how to complete the cognitive tests, after which
they spent the remaining sessions working autonomously on the assessments in two-hour
increments. I randomized each participant to work in quiet and noise for two sessions each
(see Table B13 for balance tests). Noise was generated by an engine similar to the one used
in the first experiment.
3.2 Measurement
Because there is no consensus among cognitive psychologists about the most important
measures of cognitive function and which are most likely to be relevant in this context, I
used a wide variety of tasks drawn from Dean et al. (2017), summarized in Table B12 (see
Appendix C for details). I programmed each task in an open-source, python-based platform
developed by Mathôt et al. (2012). The order of the tasks was randomly chosen for each
individual in each session. For each task, I developed a scoring rule that is a combination
of the relevant outcome measures (e.g. percentage correct and reaction time). Participants
were then paid based on their performance as measured by these scoring rules.
For analysis, I aggregate these individual test results into an index. Because the literature
thus far does not provide guidance on which aspects of cognitive function are most important
for productivity, my preferred index is the first factor of a common factor analysis of the
percentage correct and reaction times estimated using each individual’s first control session
(see Cudeck (2000) and Grice (2001) for details). This data-driven method assumes that
each measure mij of individual i on test j depends on cognitive function in the following
16linear relationship:
mij = bj ψi + Σij (3)
where ψi is the cognitive function of individual i at time t, and Σij is a noise term. The
method uses an eigenvalue decomposition to construct a set of linearly-independent factors
that approximate the measures’ covariance matrix. Assuming that all Σij are independent
of ψi and each other, any correlations between the measures can be attributed to the latent
variable ψi . Thus, the first factor, which explains the most covariance, is an index of the only
common factor ψi . For robustness, I also present the effects on the standardized total number
of points earned by a participant, the average of the standardized test scores following Kling
et al. (2007), and the first component of a principal component analysis estimated on the
same control data with similar results.
3.3 Results
3.3.1 Main Cognitive Results
As in the first experiment, treatment did not affect any environmental characteristics besides
the noise level in the room (Table 4). Moreover, the differences in average noise level between
treatment and control were also quite similar to those in experiment one (Figure 5). This
is useful because it allows me to use these results to understand the mechanisms at work in
the first experiment without strong functional form assumptions.
My preferred specification to estimate how noise affects cognitive function is the IV
using an indicator for being in a treated room as an instrument. I estimate that doubling
the perceived level of noise reduces performance on my preferred index by approximately
0.07σ (Table 5).17 This change does not appear to be driven by any particular domain
(Table B17).18 While this effect may seem small, it is important to recognize that the size of
the standard deviation is primarily driven by across-person differences (the R2 of a regression
of the index on individual fixed effects is 0.81). This implies that even substantial within-
person shifts will appear small because the measure captures size relative to differences
between individuals.
These effects are comparable to those induced by other cognitive impediments. For
example, Lichand and Mani (2016) find that a rainfall shock reduces performance on an
17
The reduced-form effect of treatment, the IV using separate instruments for different treatment intensi-
ties, and Fisher p-values presented in Table B14, Table B15 and Table B16 yield similar inferences.
18
This does not appear to be due to floor or ceiling effects, as most metrics generate good variation
(Figure A8).
17index of cognitive tests by 0.041 standard deviations. Similarly, Park (2017) finds that
a one standard deviation increase in temperature reduces students’ exam scores by 0.052
standard deviations. However, the effect is substantially smaller than the effects observed
by Mani et al. (2013), who find that once-a-year payments from sugar cane harvests increase
performance by 0.67 standard deviations.
3.3.2 Estimating the Effect of Cognitive Function on Productivity
While these effects hold independent interest, they also provide the opportunity to shed light
on the relationship between cognitive function and productivity. In particular if we believe
that noise only affects productivity through cognitive function, we can use noise exposure as
an instrument in a split-sample IV to obtain an estimate of the impact of cognitive function on
productivity (Angrist and Krueger 1992). However, we must first consider whether there are
other channels through which noise could have affected workers’ productivity. In particular,
based on the literature we should assess whether noise affected the technology of the task
directly, reduced motivation, or caused physical distress (Matthews et al. 2000b).
The first potential concern is that the noise level affects the technology of the task. For
example, if the task required coordination, the increased noise level would have likely reduced
productivity by impairing communication. As mentioned above, the task in experiment one
was chosen precisely because it does not require any kind of listening or communication
to avoid this issue. I further attempted to reduce the potential that noise could affect the
technology of the task by instructing participants in both conditions not to talk to each
other, and I randomized seat assignments to avoid participants becoming friendly with their
neighbors.
Another possible concern is that the noise level reduces a respondent’s motivation.19
To assess this possibility, participants in experiment two completed an effort task used by
DellaVigna and Pope (2018) where respondents had to alternate between pressing the “a”
and “b” keys on a keyboard for 10 minutes. The results presented in Table B18 show that
effort did not change in response to the increase in noise. The point estimate suggests that
doubling the noise level increases the number of key presses by 1.9 relative to a control
mean of 2192, and a decrease in effort larger than 1.4% is outside of the 95% confidence
interval. This is consistent with the results of the first experiment, where being in noise did
not reduce respondents’ willingness to stay and work an additional hour for a piece rate (see
Table B11).
19
There are many reasons this might be the case. For example, one might think respondents are resentful
of the noise and decide to retaliate by reducing output. Alternatively, respondents might become discouraged
by struggling to perform in noise.
18A final concern is whether the noise caused physical distress through increased stress. A
significant body of psychology literature has demonstrated that one of the ways in which an
increase in the noise level can impair cognitive function is by increasing stress levels (Szalma
and Hancock 2011). This might cause one to worry that a sufficiently large increase in
stress could have a direct effect on task performance by impairing normal body functions
like breathing and the ability to sit still. I find that stress, as measured by blood pressure,
increases as expected among individuals in the treatment condition, but that the effect is too
small to cause physical impairment (see Table B19). To put the magnitudes in perspective,
the variation is one twentieth the size that I observe in the cross-section and one twentieth
the size that Madden et al. (2017) find as residual variation after controlling for individual-
specific trends. Moreover, any physical impairment should have affected performance on the
effort task.
If cognitive function is indeed the only channel through which noise affects productivity,
the estimate of the effect on productivity from the first experiment provides a reduced-form,
and the estimate of the effect on cognitive function provides a first stage, for a split-sample
IV (Angrist and Krueger 1992). Specifically, I take the ratio of the noise level coefficients
and use the delta method to calculate standard errors. The estimates imply a substantial
“return” to temporary shifts in cognitive function (Table 6). In particular, for total pockets,
I estimate a 79% increase for every one standard deviation change in my measure of cognitive
function.20
Equipped with this estimate we can use back-of-the-envelope calculations to consider how
other cognitive impediments from the literature might influence productivity. Table 7 reports
the results of several studies examining how other stimuli can affect cognitive function. In
the last column, I multiply these effects by my estimated return to illustrate what they
might imply for productivity. For example, Ebenstein et al. (2016) find that a ten unit
increase in fine particulate matter on the day of a high-stakes exam reduces performance
by 0.017 standard deviations. Assuming that this change in performance is entirely driven
by diminished cognitive function, my estimate would imply that the same change should
reduce productivity by approximately 1.3%. Reassuringly Chang et al. (2016b) study the
productivity effect of this pollution change directly and find it reduces factory-worker output
by 8%, which suggests that, after accounting for the physical effects of pollution, an effect
of approximately 1.3% through diminished cognitive function is reasonable.
Because the impacts on cognitive function observed by Mani et al. (2013) are an order
of magnitude larger than those observed in this study, we should be cautious in applying
20
Estimates obtained using separate instruments for different treatment intensities yield larger implied
returns (Table B20).
19this estimate to their context. However, this estimate implies that the changes that they
report are likely to have significant economic impacts. This suggests that when setting
priorities policy makers should seriously consider how environments of poverty might reduce
productivity by impeding cognitive function.
4 Sorting and Efficiency
4.1 Motivation and Strategy
The combined evidence of my two experiments suggests that noise can have important im-
pacts on productivity by impeding cognitive function; however, this is not sufficient to con-
clude that the effects are relevant outside of an experimental setting. To understand whether
noise has meaningful, real-world consequences, it is vital to understand whether individuals
and firms will employ strategies to mitigate these effects. For example, the most obvious
mitigation strategy for this type of distractor is paying a compensating differential to be
able to work in quiet. If the effects observed in the experiment are driven by a subsample
who are willing to accept lower wages to work in quiet, then in equilibrium the effect might
be significantly attenuated by sorting.
Measuring individuals’ awareness of the impacts of noise also provides an opportunity to
contribute evidence on the more general question of whether individuals act strategically to
protect their cognitive function from environmental impediments. Without understanding
this level of sophistication, it’s difficult to assess the economic implications of controlled
experiments demonstrating the effects of environmental stimuli associated with poverty (see
Dean et al. (2017) for an overview). If individuals are generally aware of what situations
impair their productivity or decision-making, the substantial effects observed in controlled
experiments might be significantly attenuated by adaptation in the real-world. On the other
hand, if individuals do not understand these effects, they might cause significant inefficiencies.
For example, Schofield (2014) documents individuals failing to make food purchases that
would improve their productivity.
In order to consider the potential for adaptation, I first assess whether individuals are
aware of the impact that noise has on their productivity. At the end of each round of
both experiments, I offer participants the chance to pay for quiet working conditions and
randomly vary whether their compensation will depend on their performance.21 This allows
me to assess both individuals’ disutility from noise and the degree to which they are attuned
21
For logistical simplicity, in the second experiment individuals have their willingness to pay elicited for
a single session under the possibilities of being paid a piece rate and a flat rate. They are told that one of
their choices will be randomly implemented.
20to its impact on their productivity. In particular, workers’ willingness to pay while they
are facing a flat wage is a measure of the pure disutility value of working in noise. Any
additional amount that they are willing to pay when facing a piece rate is a measure of their
understanding of the impact that noise has on their productivity. This is because workers
who understand that noise lowers their productivity will be more willing to pay for quiet
when their pay depends on their performance. For example, the median worker in my study
can produce 13.5 perfect pockets in quiet by this point in the course. If they realize working
in noisy conditions will reduce their productivity by 3%, when they are facing a 15 Ksh piece
rate they should be willing to pay 6 Ksh more to work in quiet than when facing a pure flat
rate.
4.2 Elicitation Procedure
I elicit willingness to pay for quiet working conditions with a modified version of Becker
et al. (1964) following the approach of Berry et al. (2015) as outlined in Figure A9 using the
script in Appendix D. In this incentive-compatible task, respondents state the maximum
that they are willing to pay for a good, after which a random price is drawn. If the price is
below the respondent’s willingness to pay, he/she purchases the good at the random price,
and if the price is above the willingness to pay, the respondent does not purchase the good.
I begin with a slight modification to the procedure by employing a binary search over the
range 0-200 Ksh to identify the respondent’s maximum willingness to pay rather than begin-
ning by asking the open-ended question, “How much is the most you’re willing to pay?”.22
This procedure makes the task as concrete as possible and narrows to a final number in only
eight questions. In order to ensure understanding, after finalizing a maximum willingness to
pay, respondents must correctly answer verification questions, and they practice the entire
procedure for a lollipop. I also avoid potential issues with credit constraints, time preferences
and compliance by deducting any charges from respondents’ earnings in the session where
they paid to be in quiet. This also has the benefit of approximating a labor market situation
where individuals deciding where to work must choose whether to forgo higher potential
wages for the sake of future quiet.
22
For example, the respondent is first asked “If the random price is 100 Ksh would you want to pay to
work in quiet?” If they respond no, they are then asked “If the random price is 50 Ksh would you be willing
to pay to work in quiet?” If they respond no again, they’re asked about a random price of 25 Ksh and so on
until the search narrows to a single number.
21You can also read
